metadata
dataset_info:
features:
- name: Entry
dtype: string
- name: species
dtype: string
- name: Sequence
dtype: string
- name: labels
dtype: int64
splits:
- name: train
num_bytes: 55325670
num_examples: 147523
- name: valid
num_bytes: 1882045
num_examples: 5000
- name: test
num_bytes: 1854808
num_examples: 5000
download_size: 57511782
dataset_size: 59062523
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: valid
path: data/valid-*
- split: test
path: data/test-*
UniProt search (7/22/2025): * (Swiss-Prot reviewed entries only) download TSV with Taxonomic lineage (Ids), sequence, and length.
Curation steps:
- Filtered for sequence length >=20 and <=2048
- Extract Taxonomic IDs (domain, kingdom, phylum, class, order, family, genus, species) from taxonomic_lineage_ids column
- Keep entry, species, and sequence information
- Drop N/A
- CD-HIT using Docker at 80% similarity threshold and n=5
- Keep representative sequences
- Create labels based on spcecies
- Remove species/examples with fewer than 100 samples
- Stratified split: first get test set (5,000), then eval (5,000), rest is train